AI helps add 10k more photos to OldNYC

The update in a nutshell

It has been reported that OldNYC, the New York Public Library–sourced historic photo viewer, quietly added roughly 10,000 images over the past two years, bringing the site from about 39,000 photos in 2016 to roughly 49,000 today. The rebuild focused on three big wins: better geolocation, dramatically improved OCR, and a migration to an open mapping stack that’s cheaper and easier to run. Most of the heavy lifting happened in 2024, it has been reported, though the author only published the changes in 2026.

How AI and open data did the heavy lifting

OldNYC geocodes images by parsing textual descriptions — “Broad Street, south from Wall Street” — into coordinates. It has been reported that the developer began using OpenAI’s gpt-4o to extract detailed location information from image descriptions (not just titles), allowing the site to place photos tied to vanished buildings or renamed streets. GPT reportedly located about 6,000 additional photos, lifting the site’s locatability to roughly 87% and producing about 96% accuracy where images are mapped. OCR also got an upgrade: a custom Ocropus pipeline from 2015 gave way to modern models (gpt-4o-mini), and user edits—once a crowd-sourced “fix typos” feature—remain part of the workflow.

Map freedom and why this matters

It has been reported that OldNYC dropped Google’s geocoder in favor of OpenStreetMap, historical street datasets (including NYPL’s historical streets project), vector tiles and MapLibre — partly to avoid rising Google Maps costs and partly to remove modern anachronisms like highways that didn’t exist in the 1930s. The result: photos land where they belong, and the site runs for a hobby-friendly budget instead of a recurring API bill. Beyond dollars, there’s an emotional payoff: long-missing corners of the city reappear on the map, a kind of public-history treasure hunt powered by AI and open data. Who doesn’t love finding a buried piece of the city?

Sources: danvk.org, Hacker News